Where AI should speed up research

AI systems now produce research results faster than people can, and researchers’ responses have differed by field. Structural biologists welcomed AlphaFold’s predicted protein structures. When AI systems began settling open problems in mathematics in 2026, many mathematicians described a loss. We ask why, and where AI should be used in research. We argue that when AI produces a result, no person has produced it, so the checking, understanding and training that used to accompany production now happen only if someone undertakes them separately. Where results are wanted for use, as with a protein structure or a treatment, faster production is better, up to the capacity to check them. Mathematicians also value being the first to find a result, understanding it and the expertise formed by the work, and a faster supply of theorems provides none of these. The cost we consider most important, and least well measured, is common to all fields: the people who can check machine output learned to do so by doing work that is now being automated. We recommend that AI be used first where results are wanted for use and are cheap to check; that those who release machine-produced results report what was checked and pay for independent expert checking; that fields give credit for exposition and checking as well as for being first; and that each field keep students doing the work that forms its experts until other training is shown to work. We support the argument with documented cases from mathematics, structural biology, drug development and philosophy, and with an exploratory map of 236 research subfields. In a planned test on those subfields, growth in AI use up to 2024 was not predicted by how cheaply a subfield’s results can be checked or by how much other sciences use them.